what reward learning gets from best-of-n preference data
analysis reveals how bradley-terry reward models trained on best-of-n preference data converge to specific targets depending on n and the base distribution.
topic
research
llms learn to lie consistently across models
a multi-model study shows that fine-tuned deceptive language models develop early, linearly detectable representations of dishonesty.
python asgi apps run in browser with pyodide and service workers
a project runs python asgi web apps entirely in the browser using pyodide and a service worker, removing the need for a backend server except for static files.
momentum dynamics under sparse updates in high dimensions
new theory reveals when momentum helps or hurts in sparse training settings, based on two key timescales.
developers refuse to code without ai tools
a metr study found most developers won't work without ai, but research suggests ai may slow them down and increase maintenance costs.
ai terms you pretended to know, now explained
a plain-language glossary of common ai terms like llm, rag, and hallucination for anyone who nodded along but wants clarity.
deep learning for bivariate survival treatment rules
a new method uses deep neural networks and adaptive prediction-powered learning to optimize treatment rules for bivariate survival outcomes in randomized trials.
llm reviews show limited alignment with human judgment
a study of acl rolling review papers finds llm reviews align only moderately with human reviews and can be gamed by authors using iterative revision.
category theory boosts language model perplexity
a 306m-parameter transformer with simplicial message passing reduces perplexity by 12% over gpt-2 small on wikitext-103.
multi-task ppi boosts inference with few labels
a new framework uses labeled data from related tasks to improve statistical power in prediction-powered inference when only a handful of labels are available per task.
one mask to rule them all: hidden facts after editing
a compact binary mask reverses most knowledge edits in language models, revealing a shared mechanism behind diverse factual updates.
behavior-induced mirror-prox td for faster off-policy learning
a new mirror-prox temporal-difference method uses behavior-policy transition information instead of feature covariance to speed up off-policy prediction.
federated probe-logit distillation rate limits under heterogeneous bandwidth
new lower bounds show that the bandwidth term in federated probe-logit distillation is tight, and the method extends to nodes with different upload budgets.
ai agent plans molecular tweaks for drug lead optimization
a new llm-based agent called trace uses tool planning to optimize drug-like molecules over multiple steps, improving properties while keeping key structures intact.
behavior-aware corrections stabilize off-policy td learning
replacing the auxiliary covariance matrix with the behavior bellman matrix improves stability in off-policy temporal-difference learning.
anytime-valid federated conformal rag for llm swarms
a new method extends federated conformal rag to provide valid uncertainty estimates at any stopping time, enabling safer adaptive control in distributed language model systems.
frontier models score below 50% on enterprise it agent benchmark
itbench-aa evaluates ai agents on kubernetes incident response, with claude opus 4.7 leading at 47% accuracy.
google research at i/o 2026: ai for science, health, and more
google research highlights from i/o 2026 include new ai tools for scientific discovery, health coaching, and edge computing, plus advances in weather prediction and model factuality.